Summary
This paper (arXiv 2505.07230) introduces CapVector, a novel finetuning approach for pretrained vision-language-action (VLA) models. Standard supervised finetuning (SFT) often fails to meaningfully improve performance while keeping adaptation costs high. Advanced finetuning methods that add auxiliary training objectives can boost performance and reduce convergence steps, but they introduce significant computational overhead due to the extra losses. CapVector aims to achieve both enhanced performance and reduced adaptation cost without that overhead by learning transferable capability vectors in parametric space. The work falls in the computer vision domain, authored by Wenxuan Song, Han Zhao, and Fuhao Li, released May 9, 2025.
Paper Overview
Research Area: Computer Vision (CV)
Authors: Wenxuan Song, Han Zhao, Fuhao Li
Published: 2025-05-09
arXiv: 2505.07230
Abstract
This paper proposes a novel approach to address the challenge that pretrained VLA models often fail to effectively improve performance and reduce adaptation costs during standard supervised finetuning (SFT). Some advanced finetuning methods with auxiliary training objectives can improve performance and reduce the number of convergence steps. However, they typically incur significant computational overhead due to the additional losses from auxiliary objectives. To simultaneously achieve the enhancement of finetuning performance and the reduction of adaptation costs, the authors propose CapVector, which learns transferable capability vectors in parametric space.
Links
- arXiv page: https://arxiv.org/abs/2505.07230
---
*Auto-collected forum repost from zhichai.net, originally shared on 2026-05-13.*
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/177619927